Lesson 1 — How much maths you need
This question stops more people from starting than any technical obstacle, and it is usually answered badly, in one of two unhelpful directions. Either "you need a maths degree", which is false and discourages capable people. Or "you need no maths at all, just call the library", which is comfortable and leaves you unable to diagnose anything.
Here is a more useful answer, in three levels.
Three levels of depth
Level 1 — Using models
What you need: almost nothing formal. Understand that a model takes numbers in and produces numbers out, that "accuracy" means the proportion it got right, and that a probability near 0.5 means the model is unsure.
What you can do: train scikit-learn models on clean data, follow tutorials, produce results that are often correct.
Where it breaks: the moment something goes wrong. A model that will not converge, a metric that looks impossible, a result you need to defend in a meeting. At this level you cannot tell the difference between a bug, a data problem and a fundamental limitation, so you guess.
Level 2 — Understanding models
What you need: the intuition this course builds. Vectors as points in space, matrices as transformations, derivatives as slopes, gradient descent as walking downhill, probability as the language of uncertainty, and enough statistics to know when a difference is noise.
What you can do: diagnose why training fails, choose a loss function deliberately, know which model suits which data shape, read most of a paper, judge whether a claim transfers to your problem, and explain your results to someone who will challenge them.
This is the level nearly every AI job requires. It is the target of this course.
Level 3 — Creating methods
What you need: genuine fluency. Real analysis, optimisation theory, measure-theoretic probability, information theory, the ability to construct and follow proofs.
What you can do: invent architectures, prove properties, publish.
Who needs it: researchers. A small fraction of people working in AI, doing work that is genuinely different from applied practice.
Most intimidating advice about maths for AI is written by people at level 3, describing level 3. It is accurate for them and misleading for you. Aim for level 2 and you will be equipped for the overwhelming majority of real work.
Why school mathematics is a poor predictor
People who struggled with maths at school often assume the door is closed. It usually is not, because school mathematics and the mathematics of machine learning ask for different things.
School asked you to compute by hand: invert a matrix on paper, differentiate a complicated expression, solve for x. That is mechanical work, it is where most of the pain and most of the marks lived, and it is exactly the part a computer now does perfectly.
Machine learning asks you to understand what an operation does. What does multiplying by this matrix do to my data? Why does this derivative point in the direction that reduces the error? What does a probability of 0.7 actually commit me to?
That is a different skill, and it is more visual than procedural. Many people who found school maths dull find this version considerably more interesting, because every operation has a purpose you can see.
The three branches, and what each buys you
| Branch | What it gives you | Concretely used for |
|---|---|---|
| Linear algebra | data and transformations as geometry | everything: layers, embeddings, attention, PCA, similarity search |
| Calculus | how to improve by following a slope | training itself, backpropagation, learning rates, optimisers |
| Probability | reasoning about uncertainty | classifier outputs, loss functions, Bayesian methods, generative models |
| Statistics | telling signal from noise | evaluation, A/B testing, sampling bias, whether a result is real |
They are not equally urgent. Linear algebra is the language, so it comes first. Calculus explains training, so it comes second. Probability and statistics improve your judgement more than either, which is why practitioners who skip them keep drawing confident conclusions from noise.
What you can genuinely skip
Being explicit about this saves months of misdirected study. For applied work you can put aside, indefinitely:
- Proofs of convergence and the conditions under which an optimiser is guaranteed to work
- Measure-theoretic probability, the rigorous foundations of the subject
- Most of functional analysis and infinite-dimensional spaces
- Manual computation of anything: inverting matrices, computing determinants, symbolic differentiation
- Optimisation theory beyond knowing what the common optimisers do differently
None of these appear in a normal week of applied machine learning.
The honest cost
If you are starting from secondary-school level and want to reach level 2 properly, the realistic investment is two to four months of part-time study, which is comfortably done alongside learning Python and scikit-learn. Interleaving them is better than sequencing them: seeing a gradient in code makes the derivative concrete, and knowing what a derivative is makes the code stop looking arbitrary.
Do not study all the maths and then start. Learn a little, apply it immediately, and return for more when something confuses you. Confusion is a much better guide to what you need next than any syllabus.
In three sentences
There are three levels of mathematical depth in AI, and virtually every job needs the middle one: intuition for what operations do, not fluency in proofs. School maths is a poor predictor because it measured hand computation, which computers now do, rather than geometric understanding, which is what matters here. Two to four months of part-time study interleaved with coding gets you there, and confusion while building is a better guide than any syllabus.
Next — Lesson 2: linear algebra →